yet again, trying something new
My friend Wesley, who I met in Amsterdam at Thom’s 50th birthday, has a newsletter called Process and I’ve been inspired by how he thinks and works.
So I’m just going to talk about some things that are in process or on my mind and jot things down as they come to mind throughout the week.
Here’s what came out.
Art vs Decoration
Walking around the Minneapolis Institute of Art with Abraham, looking at a large mural made out of driftwood.
“This… does nothing for me. It’s pretty, but I don’t feel it,” I said.
“Yeah, I always ask myself the question: is this art or is this decoration?” Abraham replied.
As we walked around we started calling it out.
A skeleton hanging around a torch: “Art.”
A carpet with a cool design: “Decoration.”
We arrived at a large abstract piece, both silent for a few seconds as we took it in.
We looked at each other and said at the same time:
“Neither.”
Railway makes cloud computers awesome
I’ve really been enjoying Railway as a cloud server for Edgar.
Edgar is my virtual butler / project manager that I’ve been building in my spare time or when I get inspired. It’s important for me to have outlets like this because it helps me learn new tech which usually inspires a breakthrough at work.
Edgar updates my wiki, takes meeting notes, organizes everything into tasks and next actions like a project manager, does research, schedules things on my calendar, and is just sort of an always-there assistant.
He connects to Notion and Slack and I can send him messages by holding down the action button on my phone … honestly this guy is saving my life.
I created him in Copenhagen for some of our investment meetings, where we needed to do a bunch of things at once and it was just way too much to keep in our heads. So now Edgar listens and creates and manages everything and I really want to do it in a way that as few companies have access to the data as possible (which is why I use a local model for transcription and then go direct to Anthropic for anything else).
Also for a really practical reason: I don’t want to be in those meetings, but I do want a feedback loop and to be able to see where tasks are at any given time in an organized, updated way.
Edgar is really good at this. Way better than me even if I were to be attending the meetings and taking notes and updating tasks. It’s sort of like having a 24/7 project manager who keeps everything moving.
Anyway, to set him up I needed a cloud server, and I hadn’t set up infrastructure in quite a few years so I haven’t used some of the latest cool new tech.
It was so cool to use Railway. I can’t believe how cheap and easy it is. They have an agent in there which I used to deploy everything. I connected a GitHub repository where anytime something is pushed to it, it automatically deploys to the server. Then in between connections the server goes to sleep, so it costs like a penny a day — it’s so ridiculously cheap. It’s also fun to use, and you can see all the logs. This is my cost for over a week of usage… $0.05:
You can put all of your secrets, like API keys, into variables so it’s a lot more secure than what I was doing before (being bad and storing API tokens in a Claude Cowork file but it was a good proof of concept).
Anyway I’m just obsessed with all of this, and it makes me feel like anything’s possible.
Glendale to Altadena on an e-scooter
I tried riding my e-scooter from Glendale to Altadena.
I've really been obsessed with biking ever since we spent time in Amsterdam and Copenhagen. In Los Angeles we spend so much time in our car — hours every day. It's completely ridiculous. So I just wanted to see if I could get to Betsy Bar, which now runs brunch, on a scooter from Glendale.
I charged it up, because I hadn't used it in about a year — I have a Segway Ninebot — and set off. After about 50 minutes of scootering I ALMOST made it there, probably 95% of the way, and then my battery died.
There were some nice stretches, especially through Chevy Chase and Flintridge. There were also stretches that were genuinely scar with huge roads and huge cars leaving me feeling extremely exposed.
I thought I could walk there since I was close but directions said it was still a 35 minute walk and I had to push the scooter so I called Emily and she rescued me, dropped me off, and kept the scooter so we could charge it later.
Conclusion: Despite it not being super inspiring, I’m going to try it again, but on a bike. I rented some bikes so we could give it a try as a family. In the meantime I’m going to be dreaming about Copenhagen.
And so is Zizi (you have to watch this video of her narrating while in the front of a cargo bike in Copenhagen).
Finally, a compliment from Edgar
I’ve started creating multiple instances of Edgar for different purposes, hard segmented with different security privileges.
It’s gone from a tool I had to communicate with chat to where I can hold the action button down on my phone and now it sends an audio file directly to Edgar … no need to open up any app at all or use the screen.
Edgar processes it, puts it in Slack (nicely threaded!), and then updates all the tasks/wiki/notes in Notion. Then I can respond to him right in Slack as well. So now everything is threaded, organized, and maintainable, all from a one-second button click on my phone.
It’s a really thrilling feeling.
But anyway, out of nowhere Edgar gave me this compliment, which was the first time he ever said anything positive (he’s a bit of a grumpy butler):
Go OFF my butler friend (oh no, am I going into AI psychosis?).
Every now and then I go through periods of life where I don’t know if I’m ever going to build again. Like… have I become a dry well? Am I too old and tired? Am I falling behind on the latest tech? But it always comes back. I always build something, and I’m always surprised at how life-giving it is.
My curly hair needs curly hair help … and maybe softer water?
Emily surprised me with a new haircut place that specializes in curly hair. I don't normally think of my hair as curly — I think of it as a messy mop that I never touch.
I somewhat reluctantly followed my calendar to an address at an unmarked building, where a 23-year-old with the most LA accent I have ever heard in my life taught me all about the things I have never thought about and would never do to my hair.
Even without any product it made a huge difference in how my hair sat.
She said things like: Put your head toward the floor and shake it out like you're a dog. Okay, I can do that. Use a t-shirt instead of a towel to dry it, and clump it instead of tornado-ing it. Okay, I can do that too — although that one is a little more annoying. Now I need a towel and a t-shirt next to the shower?
The other one was to get a shower filter to reduce the hardness of the water. This one I was skeptical of. But she was right about the other things, so I put Edgar on the task and asked him to research whether these actually work and what to get if so.
I'd give him a C. He said shower filters don’t really work, but then still suggested ones to buy. Like, bruh, don’t do that. So I'm tweaking him to be more skeptical and opinionated on these sorts of things (after researching — I also need to tweak the kinds of things he believes).
BUT! He did surface a shampoo that removes the minerals from my hair, which hopefully will work. He linked to it on Amazon, I ordered it, and he marked the task off.
Honestly this felt super successful — normally I would have totally forgotten to do this, or research it, and it would have gotten lost. What I'd really like is the ability for him to just order something on Amazon, but I haven't figured that one out yet in a safe way.
LED product catalog as kid’s toy
I got an LED lighting product catalog for my kid to play with.
A month or so ago I was down some rabbit hole and saw an LED light catalog that I was floored by. It was like 10 pages, but each page was an actual circuit board with multiple LEDs on it. You plugged in a USB-C and could cycle through them. It was so cool and such a great way to make a portable way to show off your wares.
A couple years ago Emily and I had a meeting with a money guy who brought these folders that had an Android tablet in them … when you opened it, it would auto-play the video of what they wanted you to invest in. That was very funny to me and we tried to keep them. But they pestered us to send them back, which I thought was a great representation of why not to invest with them. If you don’t let people keep your sales material, I'm not going to invest money with you.
Anyway, back to the light book: I was able to track down the supplier, contact them on WhatsApp, and get them to send one. Because of shipping and tariffs and all that stuff, it ended up being more than I had expected for, you know, electronic sales literature.
The other day I got it in and it was exactly what I was hoping for. And I was curious to see if Zizi would like it. I mean, I thought it was cool, but I really got it for her.
It was a hit. She loved playing with it. And Emily had a great idea for her to wear sunglasses while she turned the brightness up
She was a bit rough with it so when I told her to be careful, she told me she could just buy another one.
“How many dollars do you have?” I asked.
“D’eleven,” she said.
“Well, that light book is like two hundred dollars.”
“WoWooow,” she said, “you should really save your money.”
Hard to argue with that.
I'm still laughing at the memory of her in the backseat wearing her sunglasses, playing with the bright lights.
Now I can just ask Edgar to do anything from anywhere
My very first version of Edgar lived entirely in Claude Cowork as a bunch of markdown instruction files. I did this so that he ran on a “free” cloud container that would run from my phone or desktop and would keep running when I shut my laptop. It was a hack.
My second version was a single Edgar on a Railway server where I was able to send voice notes from my phone, which then posted the reply in Slack and made changes to tasks and wiki in Notion.
My third version enabled two-way communication from Slack.
My fourth version created mini-Edgars with sandboxed credentials and filesystems so that I could use Edgar in more secure contexts than my personal task list.
The two-way communication was a big unlock. For instance, one of the first things I did was ask what things we wanted to see when my mom was visiting me in California in a couple weeks. And he pulled it out of a task. Handy:
(I did make some style and tone changes based on this response, which I tweak a bit each day still.)
This is allowing me to build a wiki on the go that doubles as a task management system … then I can just ask questions of it! Which is making the actual filing system itself increasingly irrelevant. There needs to be architecture behind it, which I'm providing, but the retrieval of it will increasingly be just natural conversation and questions.
I’m able to do that inside an interface I'm very comfortable with (Slack). And now it's organized, threaded, searchable, and inside my own system.
Now I’m wondering if I should record everything 24/7??
Bring your Zizi to work day
This week it was Bring Your Kid to Work Day. So we stayed at the hotel by the office in Playa del Rey and met up for dinner with my boss and his daughter.
It's so funny to me that I say I have a boss. I don't really know what that means. Like, what does it mean to have a boss? I haven't had a boss in 20 years. And I do now, and amazingly it works. Thank goodness for all parties involved.
Maybe a boss is someone you have to ask permission for to post a picture. You know what, I’m just going to blur instead of ask permission because honestly that’s just not who I am.
LET ME KNOW IF YOU WANT ME TO UNBLUR THIS, BOSSSSSSSSSS!!!
Well, we met up for dinner at our favorite spot in Santa Monica. Close enough to smell the ocean air and still walk to ice cream.
In the car on the way, we had practiced some of the questions that Zizi would be asked in front of everyone at the Town Hall Meeting the next day.
The first question was:
what do your parents do when no one's looking?
Emily and I looked at each other with raised eyebrows. Weird question, what would she say? She said: “uhhhhhhHHhhhh I don't know.” We explained that that was not an interesting answer.
“Business, I guess?” she responded.
Well, the answer could be worse.
Still, we suggested some revision to the questions.
It went really well! Here’s the video:
You can hear the murmuring from everyone after her answer to “where is your favorite place to go with your mom?” The other kids said the park, the beach… Zizi straight up said “Copenhagen.” We were laughing. Someone asked us if we were from Denmark, which would make that answer make more sense, but we just shook our heads no, laughing.
Love how not boring she is.
Making Edgar faster for no reason
There’s really not any reason to make Edgar faster. He's fast enough. What am I going to do, NOT wait 60 seconds for a response?
Most of the time I don't even need a fast-ish response — it's just stuff I'm sending that I can look at later, or that I never need to look at at all because it's filed into the wiki. But I can't stop myself. I want to make things faster. It's fun and it’s easy to measure.
So I had Edgar measure himself and timed how long each part was taking:
Naughty naughty Edgar, those synchronous calls are bad! He was grabbing his own instructions every time, and then pulling all of the wiki data every time and processing it, instead of pulling everything that was already stored once and then only doing what had been edited.
I gave him a goal for the time (both humans and agent loops thrive with clear goals) and after some back-and-forths and testing we got the average response time down from 60s to 20s.
Here’s the weird part: I dictated this to Edgar to file in this newsletter so I didn’t know where we’d actually arrive at the final response times. So in the most meta fashion, this is the first real test of the new architecture!
Don’t bring a pointy stick to the art museum
At the museum there was a man walking around holding a long wooden skewer with a sharp end.
Something seemed a bit off, and that became really apparent when he started dragging the skewer across artwork.
“Sir, please don't do that,” one of the minders said. In silent reply, the man walked to the next painting and dragged it across that one.
“Sir,” … security started running towards him … “you can't keep poking the art repeatedly with a sharp stick!”
The pokey stick guy shuffled fast to the next room and continued.
We watched as people tried to figure out what to do.
Some of that art might be worth slightly less now.
Or maybe more?
Super Edgars and Mini Edgars
As soon as I showed Emily what I was doing with Edgar, she really wanted access, which was really motivating for me. After a week or so she was literally begging for an Edgar … which meant figuring out how to do that.
I wanted to start out with a limited scope for household management stuff. But it's really important that a mini Edgar doesn't have access to absolutely everything. I don't want things getting mixed up — like things going into my personal wiki or tasks when we're trying to get them into the household management area. So these structures need to be very separate.
I spent some time wireframing some potential solutions at a coffee shop:
I fed it into Claude Fable to create diagrams for me and to simplify it further which turned out really well (after, of course, many rounds of back and forth). Here’s an example early diagram output:
Originally it had proposed a solution which was super convoluted (encrypting the other files and folders each time an instance ran??!!) and it hardly took any pushback at all for it to say, yeah, that was a terrible idea, let’s not do that.
This is the weakness of these kinds of models: even the smartest models have a hard time creating simple architecture. I'm not an infrastructure whiz, but I know enough to know that was a dumb solution.
I suggested we use a more native Unix-based approach where we have a user account for each of the mini Edgars on the same server. Then the data is partitioned by the operating system instead of by some weird encryption scheme that a blind hamster would be able to crack.
Of course these models are all insanely sycophantic, and so it said, oh yeah, that's a much better method. Since we both thought it was a good idea, we built it like that. So now I'm rebuilding the architecture. And fingers crossed that Edgar still works.
Update: It did and now Edgar has babies.